Segmentation of Rumex obtusifolius using Gaussian Markov random fields

Abstract

Rumex obtusifolius is a common weed that is difficult to control. The most common way to control weeds—using herbicides—is being reconsidered because of its adverse environmental impact. Robotic systems are regarded as a viable non-chemical alternative for treating R. obtusifolius and also other weeds. Among the existing systems for weed control, only a few are applicable in real-time and operate in a controlled environment. In this study, we develop a new algorithm for segmentation of R. obtusifolius using texture features based on Markov random fields that works in real-time under natural lighting conditions. We show its performance by comparing it with an existing real-time algorithm that uses spectral power as texture feature. We show that the new algorithm is not only accurate with detection rate of 97.8 % and average error of 56 mm in estimating the location of the tap-root of the plant, but is also fast taking just 0.18 s to process an image of size \(576 \times 432\) pixels making it feasible for real-time applications.